⭐️ Ai Misuse & Intune Enrollments

September 17, 2025 – Lab Notes

Research & Exploration

  • Asked about Claude AI misuse and adversarial fine-tuning risks.
    • Key takeaway: LoRA/PEFT can be abused to alter model behavior.
    • Learned that poisoned fine-tuning data can act as a backdoor, enabling unauthorized actions.
    • Example risks: data exfiltration, lateral movement, adversarial prompts.
  • Explored open weights:
    • Parameters are trained weights.
    • “Open weights” = public release of those trained parameters.
    • These can be wrapped into an agent and run locally.
  • Learned about mixture-of-experts (MoE):
    • Technique where only a subset of model experts are activated per query → efficiency and scaling.
  • Built an LLM + Agent + Tools diagram to visualize roles:
    • LLM = core reasoning engine.
    • Agent = orchestration + decision layer.
    • Tools/Environment = APIs, shells, browsers, etc.

diagram

Sys Admin

  • Investigated Intune enrollments.